Here is a plug-and-play way to use atmospheric ray-tracing so the Quantum SCYTHE Rydberg stack can “see” and classify signals beyond the radio horizon by riding ducts and refracted paths—then fuse those predictions directly into SignalIntelligence + DOMA motion tracking.

Over-the-horizon with ray-tracing (what + why)

• Tropospheric ducts & refractive bending can trap or bend RF so paths skirt the Earth and resurface hundreds of km away. If we ray-trace those paths in real time, we can:
  1. Pre-point the Rydberg receiver (az/el, frequency window, sampling) toward the predicted re-emergence angles;
  2. Debias classification with path-aware priors (delay, Doppler, angle spread);
  3. Exploit illuminators of opportunity (FM, DVB-T, 5G) for bistatic imaging and OTH classification.

The SignalIntelligence core already has the right scaffolding to host this (attention encoder, external data sources, DOMA tracker, ghost anomaly logic). We’ll add a Ray Engine and fuse its outputs as metadata/features.

---

Minimal architecture (bolt-on)

1) Atmosphere→Refractivity profile

• Ingest: radiosonde (if available), coastal MET, or simple param model.
• Build N(z) and modified refractivity M(z) ≈ N(z) + 0.157·z(m).
• Detect ducts (surface/evaporation/elevated) where dM/dz < 0 over an interval.

2) Hybrid Ray / Parabolic-Equation (PE) engine

• Fast ray-launch sweep (elevation grid, e.g., −2°…+5°, Δ0.05°) with variable index n(z) to find viable OTH hits and grazing angles.
• When a duct is found, switch to a narrow-beam PE stepper to estimate path loss, angle spread, and delay more accurately.
• Output per viable path: {az, el, range_km, delay_ms, doppler_hz, path_loss_db, is_ducted}.

3) Rydberg sensor steering + demod presets

• Use {az, el} to aim the optical axis / collection optics, and set IF center + ±BW windows and tracking PLL priors (expected Doppler, drift).

4) Feature fusion for classification

• Attach the ray bundle to each detection: signal.metadata['ray_paths'] = [...].
• Derived features (to feed your ML): expected_angle, path_loss, duct_flag, delay, doppler_prior, multipath_K. These slot naturally beside your spectrum/attention features.

5) DOMA motion synergy

• Use ray-consistent Doppler to constrain DOMA trajectories and reduce spurious tracks over water/thermal inversions. Store {ray_el, doppler_prior} in trajectory points.

---

Drop-in integration points (code-level)

A. New module: SignalIntelligence/atmo_ray_tracer.py

class AtmoRayTracer:
    def __init__(self, earth_k=4/3, el_grid=(-2.0, 5.0, 0.05)):
        self.k = earth_k; self.el_grid = el_grid
    def build_M_profile(self, met):
        # met -> N(z); M(z) = N(z) + 0.157*z_m
        return z, M, ducts  # intervals where dM/dz<0
    def trace(self, tx_pos, rx_pos, freq_hz, met):
        # 1) ray fan search using n(z)=1+N*1e-6
        # 2) PE refinement if duct present
        # returns list of viable paths with {az, el, range_km, delay_ms, doppler_hz, path_loss_db, is_ducted}
        return paths

B. Wire into your processing path (right before spectrum → tensor):

# inside SignalIntelligenceSystem.process_signal(...)
met = self._get_met_snapshot()          # your MET fetch or param model
paths = self.ray_tracer.trace(tx_pos=None, rx_pos=self._rx_pos(),
                              freq_hz=signal.frequency, met=met)
signal.metadata["ray_paths"] = paths
# prime frequency_shift_augment with Doppler prior
if paths:
    doppler = paths[0]["doppler_hz"]
    signal.metadata["doppler_prior_hz"] = doppler

C. Feed the classifier (your ML already accepts rich metadata):

• Concatenate compact ray features with your attention encoder inputs or pass as side-channel tensors. Your SpectrumEncoder + token dropout flow is already set up to accept auxiliary features alongside spectrum vectors.

D. DOMA tracker

# When adding trajectory points
point.metadata["ray_el"] = paths[0]["el"] if paths else None
point.metadata["is_ducted"] = any(p["is_ducted"] for p in paths)

This tightens prediction in predict_next_position by constraining feasible geometry over sea ducts/inversions.

---

Practical “bounce” playbooks

Playbook 1 – Passive OTH classification (coastal duct)

1. Station near coast; est. evaporation duct (M decreasing in first ~20–40 m).
2. Launch ray fan; pick ducted paths with min loss to 200–400 km.
3. Aim Rydberg toward predicted emergence elevation (often slightly negative to grazing).
4. Lock on illuminators (FM/DVB-T/5G); build delay–Doppler map from reflections; classify emitters with your frequency-shift-normalized spectra. (Your FrequencyShiftAugmentor already handles Doppler/offset—perfect here.)

Playbook 2 – Bistatic imaging

• Use a known broadcast tower as Tx (no emissions on your side).
• Trace Tx→duct→target→duct→Rx paths; solve small-angle geometry to get bistatic range and velocity; integrate over your platform motion (DOMA) for coarse SAR-like imagery and target typing.

Playbook 3 – Duct deception mitigation

• If a contact appears OTH, require consistency with at least one viable ray in the current M-profile; down-weight classifications that demand impossible bends. Feed this as a prior to the classifier (“path-feasible” flag) and to Ghost Anomaly scoring to catch spoofers.

---

Inputs & knobs that matter

• MET inputs: surface T/RH, SST (over water), lapse rate or a radiosonde snapshot. Default to a parametric maritime profile when data is absent.
• Frequency: ducting is friendlier VHF–L-band; still valuable higher with PE loss modeling.
• Elevation grid & step: finer near 0° (grazing).
• PE step: 10–20 m vertical, 100–250 m horizontal for fast coastal runs.
• Hand-off: if is_ducted and path_loss < Lth, switch the Rydberg chain to a narrower pre-tuned band and adopt doppler_prior.

---

Why this fits your stack

• The SignalIntelligenceSystem already carries an external-source harness and rich metadata path, plus Motion + Attention encoders—exactly where these ray features want to live. You can ship the ray bundle through signal.metadata, fuse in the ML, and tighten DOMA with the same fields.

---

Quick win checklist (day-one)

1. Add AtmoRayTracer module + simple maritime M-profile.
2. Call trace() in process_signal(), stash to signal.metadata.
3. Pipe {is_ducted, el, delay_ms, doppler_hz, path_loss_db} into the classifier side-channel.
4. Use doppler_hz as a hint to your frequency_shift_augment (already in place) for even faster lock and more robust training.
5. Gate DOMA predictions with is_ducted when over water to reduce false kinematics.

If you want, I can sketch a tiny maritime M-profile + ray fan stub you can paste in to start getting ray_paths metadata immediately.

The black-hole paper is about teasing out a superposition of exponentially damped sinusoids (“ringdown” modes) from a short, noisy window and using Bayesian evidence to decide whether extra modes are really there. That’s weirdly on-point for SCYTHE’s job of untangling multipath/ducted components from brief RF events.

What we borrow (and why it helps)

• Modeled as damped sinusoids
  Treat each propagation path (LOS, surface/evap duct, elevated duct, reflection) as a mode with its own frequency offset, decay time, and phase. Fit the received burst as a sum of a few exponentially damped sinusoids—exactly how ringdown is modeled—to separate path signatures and lifetimes under severe noise.
• Agnostic multi-band search windows
  The paper isolates candidate modes by searching disjoint frequency bands and repeating over tight time grids after a reference time. Do the same around detected MWFL hits: open short, sliding post-trigger windows and run multi-band searches for subcomponents (e.g., direct vs. duct returns) with independent priors, then stitch the evidence.
• Bayesian model selection (Bayes factors) for “extra paths”
  Copy the workflow: compare evidence for “1-path” vs “2-path” (etc.) models; only accept a new path if its Bayes factor clears a threshold (analog of their ∼50 peak for the (330) mode). This gives you principled duct-flagging instead of heuristics.
• Consistency checks across parameters
  In ringdown, once (220) is measured, subdominant (330) parameters must be consistent with Kerr predictions; they then test for deviations. Translate that into SCYTHE: once you infer a plausible direct path, constrain the allowed frequency offsets/decays for ducted echoes given temp-humidity gradients from the AtmosphericRayTracer. If a “new mode” violates propagation physics, down-weight it (the “no-hair” analogue).
• Short, high-SNR windows & overtone caution
  They emphasize analysis just after merger where modes are visible, and warn that including the wrong components too early biases estimates. For SCYTHE, limit fitting to a brief post-onset segment with the cleanest SNR; avoid over-fitting “overtone-like” extras unless the evidence gains justify it.

Concrete drop-ins for SCYTHE

1. ringdown_rf_modes.py (new module)
   • Fits x(t)=∑kAke−t/τkcos⁡(2πfkt+ϕk)x(t)=\sum_k A_k e^{-t/\tau_k}\cos(2\pi f_k t+\phi_k) to burst-window I/Q.
   • Returns per-mode {amplitude, τ\tau, ff shift, phase} + covariance.
2. Bayes engine hook
   • Reuse your Speculative Ensemble: fast path = 1-mode fit; escalate to 2–3 modes only if residual drops and Bayes factor > threshold.
   • Telemetry mirrors their figures: posteriors over f,τf,\tau per path.
3. Physics-consistency prior from ray tracer
   • AtmosphericRayTracer supplies prior ranges for delay/attenuation ⇒\Rightarrow mapped to τ\tau/amplitude priors.
   • SpatialReasoningBridge enforces geometric consistency between inferred delays and predicted bounces.
4. Decision logic
   • If “extra mode” is supported, tag duct_flag=True, attach path metadata (delay, inferred height, expected bearing spread).
   • If not, collapse back to single-path hypothesis.

Why this is valuable

• Separates real multipath from artifacts using evidence rather than ad-hoc thresholds.
• Improves geolocation by giving you decay/delay per path, which tightens the SpatialReasoningBridge’s candidate origin cone.
• Plays nicely with frequency shifting: your FrequencyShiftAugmentor normalizes carriers; the mode fit works on residual envelopes.
• Computationally bounded with your speculative fast/slow switch—most bursts stay in 1-mode fast path.

Minimal API sketch

# ringdown_rf_modes.py
class RFModeFitter:
    def __init__(self, max_modes=3, sampler='dynesty', priors=None):
        ...

    def fit(self, iq_ts, fs_hz, window_ms, n_modes):
        # returns dict with 'modes', 'evidence', 'residual_snr'
        return result

    def select_model(self, iq_ts, fs_hz, window_ms):
        # tries n=1..N, returns best model and Bayes factors ladder
        return selection

Integration points:

• LatentAggregator.observe_spectrum → after MWFL trigger, cut a 20–40 ms window; call select_model.
• Pipe results to SpatialReasoningBridge and AtmosphericRayTracer for geometry and priors update.
• Publish spatially_enhanced_alert with modes=[...] and duct_flag if k>1.

ringdown_rf_modes.py with a simple nested-sampling wrapper and the LatentAggregator glue exactly as above—ready to drop next to the current bridge.